| 1 | //===----- RISCVCodeGenPrepare.cpp ----------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This is a RISC-V specific version of CodeGenPrepare. |
| 10 | // It munges the code in the input function to better prepare it for |
| 11 | // SelectionDAG-based code generation. This works around limitations in it's |
| 12 | // basic-block-at-a-time approach. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #include "RISCV.h" |
| 17 | #include "RISCVTargetMachine.h" |
| 18 | #include "llvm/ADT/Statistic.h" |
| 19 | #include "llvm/Analysis/ValueTracking.h" |
| 20 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 21 | #include "llvm/IR/Dominators.h" |
| 22 | #include "llvm/IR/IRBuilder.h" |
| 23 | #include "llvm/IR/InstVisitor.h" |
| 24 | #include "llvm/IR/Intrinsics.h" |
| 25 | #include "llvm/IR/PatternMatch.h" |
| 26 | #include "llvm/InitializePasses.h" |
| 27 | #include "llvm/Pass.h" |
| 28 | #include "llvm/Transforms/Utils/Local.h" |
| 29 | |
| 30 | using namespace llvm; |
| 31 | |
| 32 | #define DEBUG_TYPE "riscv-codegenprepare" |
| 33 | #define PASS_NAME "RISC-V CodeGenPrepare" |
| 34 | |
| 35 | namespace { |
| 36 | class RISCVCodeGenPrepare : public InstVisitor<RISCVCodeGenPrepare, bool> { |
| 37 | Function &F; |
| 38 | const DataLayout *DL; |
| 39 | const DominatorTree *DT; |
| 40 | const RISCVSubtarget *ST; |
| 41 | |
| 42 | public: |
| 43 | RISCVCodeGenPrepare(Function &F, const DominatorTree *DT, |
| 44 | const RISCVSubtarget *ST) |
| 45 | : F(F), DL(&F.getDataLayout()), DT(DT), ST(ST) {} |
| 46 | bool run(); |
| 47 | bool visitInstruction(Instruction &I) { return false; } |
| 48 | bool visitAnd(BinaryOperator &BO); |
| 49 | bool visitIntrinsicInst(IntrinsicInst &I); |
| 50 | bool expandVPStrideLoad(IntrinsicInst &I); |
| 51 | bool widenVPMerge(IntrinsicInst &I); |
| 52 | }; |
| 53 | } // namespace |
| 54 | |
| 55 | namespace { |
| 56 | class RISCVCodeGenPrepareLegacyPass : public FunctionPass { |
| 57 | public: |
| 58 | static char ID; |
| 59 | |
| 60 | RISCVCodeGenPrepareLegacyPass() : FunctionPass(ID) {} |
| 61 | |
| 62 | bool runOnFunction(Function &F) override; |
| 63 | StringRef getPassName() const override { return PASS_NAME; } |
| 64 | |
| 65 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 66 | AU.setPreservesCFG(); |
| 67 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 68 | AU.addRequired<TargetPassConfig>(); |
| 69 | } |
| 70 | }; |
| 71 | } // namespace |
| 72 | |
| 73 | // Try to optimize (i64 (and (zext/sext (i32 X), C1))) if C1 has bit 31 set, |
| 74 | // but bits 63:32 are zero. If we know that bit 31 of X is 0, we can fill |
| 75 | // the upper 32 bits with ones. |
| 76 | bool RISCVCodeGenPrepare::visitAnd(BinaryOperator &BO) { |
| 77 | if (!ST->is64Bit()) |
| 78 | return false; |
| 79 | |
| 80 | if (!BO.getType()->isIntegerTy(Bitwidth: 64)) |
| 81 | return false; |
| 82 | |
| 83 | using namespace PatternMatch; |
| 84 | |
| 85 | // Left hand side should be a zext nneg. |
| 86 | Value *LHSSrc; |
| 87 | if (!match(V: BO.getOperand(i_nocapture: 0), P: m_NNegZExt(Op: m_Value(V&: LHSSrc)))) |
| 88 | return false; |
| 89 | |
| 90 | if (!LHSSrc->getType()->isIntegerTy(Bitwidth: 32)) |
| 91 | return false; |
| 92 | |
| 93 | // Right hand side should be a constant. |
| 94 | Value *RHS = BO.getOperand(i_nocapture: 1); |
| 95 | |
| 96 | auto *CI = dyn_cast<ConstantInt>(Val: RHS); |
| 97 | if (!CI) |
| 98 | return false; |
| 99 | uint64_t C = CI->getZExtValue(); |
| 100 | |
| 101 | // Look for constants that fit in 32 bits but not simm12, and can be made |
| 102 | // into simm12 by sign extending bit 31. This will allow use of ANDI. |
| 103 | // TODO: Is worth making simm32? |
| 104 | if (!isUInt<32>(x: C) || isInt<12>(x: C) || !isInt<12>(x: SignExtend64<32>(x: C))) |
| 105 | return false; |
| 106 | |
| 107 | // Sign extend the constant and replace the And operand. |
| 108 | C = SignExtend64<32>(x: C); |
| 109 | BO.setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: RHS->getType(), V: C)); |
| 110 | |
| 111 | return true; |
| 112 | } |
| 113 | |
| 114 | // With EVL tail folding, an AnyOf reduction will generate an i1 vp.merge like |
| 115 | // follows: |
| 116 | // |
| 117 | // loop: |
| 118 | // %phi = phi <vscale x 4 x i1> [ zeroinitializer, %entry ], [ %rec, %loop ] |
| 119 | // %cmp = icmp ... |
| 120 | // %rec = call <vscale x 4 x i1> @llvm.vp.merge(%cmp, i1 true, %phi, %evl) |
| 121 | // ... |
| 122 | // middle: |
| 123 | // %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %rec) |
| 124 | // |
| 125 | // However RVV doesn't have any tail undisturbed mask instructions and so we |
| 126 | // need a convoluted sequence of mask instructions to lower the i1 vp.merge: see |
| 127 | // llvm/test/CodeGen/RISCV/rvv/vpmerge-sdnode.ll. |
| 128 | // |
| 129 | // To avoid that this widens the i1 vp.merge to an i8 vp.merge, which will |
| 130 | // generate a single vmerge.vim: |
| 131 | // |
| 132 | // loop: |
| 133 | // %phi = phi <vscale x 4 x i8> [ zeroinitializer, %entry ], [ %rec, %loop ] |
| 134 | // %cmp = icmp ... |
| 135 | // %rec = call <vscale x 4 x i8> @llvm.vp.merge(%cmp, i8 true, %phi, %evl) |
| 136 | // %trunc = trunc <vscale x 4 x i8> %rec to <vscale x 4 x i1> |
| 137 | // ... |
| 138 | // middle: |
| 139 | // %res = call i1 @llvm.vector.reduce.or(<vscale x 4 x i1> %rec) |
| 140 | // |
| 141 | // The trunc will normally be sunk outside of the loop, but even if there are |
| 142 | // users inside the loop it is still profitable. |
| 143 | bool RISCVCodeGenPrepare::widenVPMerge(IntrinsicInst &II) { |
| 144 | if (!II.getType()->getScalarType()->isIntegerTy(Bitwidth: 1)) |
| 145 | return false; |
| 146 | |
| 147 | Value *Mask, *True, *PhiV, *EVL; |
| 148 | using namespace PatternMatch; |
| 149 | if (!match(V: &II, |
| 150 | P: m_Intrinsic<Intrinsic::vp_merge>(Op0: m_Value(V&: Mask), Op1: m_Value(V&: True), |
| 151 | Op2: m_Value(V&: PhiV), Op3: m_Value(V&: EVL)))) |
| 152 | return false; |
| 153 | |
| 154 | auto *Phi = dyn_cast<PHINode>(Val: PhiV); |
| 155 | if (!Phi || !Phi->hasOneUse() || Phi->getNumIncomingValues() != 2 || |
| 156 | !match(V: Phi->getIncomingValue(i: 0), P: m_Zero()) || |
| 157 | Phi->getIncomingValue(i: 1) != &II) |
| 158 | return false; |
| 159 | |
| 160 | Type *WideTy = |
| 161 | VectorType::get(ElementType: IntegerType::getInt8Ty(C&: II.getContext()), |
| 162 | EC: cast<VectorType>(Val: II.getType())->getElementCount()); |
| 163 | |
| 164 | IRBuilder<> Builder(Phi); |
| 165 | PHINode *WidePhi = Builder.CreatePHI(Ty: WideTy, NumReservedValues: 2); |
| 166 | WidePhi->addIncoming(V: ConstantAggregateZero::get(Ty: WideTy), |
| 167 | BB: Phi->getIncomingBlock(i: 0)); |
| 168 | Builder.SetInsertPoint(&II); |
| 169 | Value *WideTrue = Builder.CreateZExt(V: True, DestTy: WideTy); |
| 170 | Value *WideMerge = Builder.CreateIntrinsic(ID: Intrinsic::vp_merge, Types: {WideTy}, |
| 171 | Args: {Mask, WideTrue, WidePhi, EVL}); |
| 172 | WidePhi->addIncoming(V: WideMerge, BB: Phi->getIncomingBlock(i: 1)); |
| 173 | Value *Trunc = Builder.CreateTrunc(V: WideMerge, DestTy: II.getType()); |
| 174 | |
| 175 | II.replaceAllUsesWith(V: Trunc); |
| 176 | |
| 177 | // Break the cycle and delete the old chain. |
| 178 | Phi->setIncomingValue(i: 1, V: Phi->getIncomingValue(i: 0)); |
| 179 | llvm::RecursivelyDeleteTriviallyDeadInstructions(V: &II); |
| 180 | |
| 181 | return true; |
| 182 | } |
| 183 | |
| 184 | // LLVM vector reduction intrinsics return a scalar result, but on RISC-V vector |
| 185 | // reduction instructions write the result in the first element of a vector |
| 186 | // register. So when a reduction in a loop uses a scalar phi, we end up with |
| 187 | // unnecessary scalar moves: |
| 188 | // |
| 189 | // loop: |
| 190 | // vfmv.s.f v10, fa0 |
| 191 | // vfredosum.vs v8, v8, v10 |
| 192 | // vfmv.f.s fa0, v8 |
| 193 | // |
| 194 | // This mainly affects ordered fadd reductions and VP reductions that have a |
| 195 | // scalar start value, since other types of reduction typically use element-wise |
| 196 | // vectorisation in the loop body. This tries to vectorize any scalar phis that |
| 197 | // feed into these reductions: |
| 198 | // |
| 199 | // loop: |
| 200 | // %phi = phi <float> [ ..., %entry ], [ %acc, %loop ] |
| 201 | // %acc = call float @llvm.vector.reduce.fadd.nxv2f32(float %phi, |
| 202 | // <vscale x 2 x float> %vec) |
| 203 | // |
| 204 | // -> |
| 205 | // |
| 206 | // loop: |
| 207 | // %phi = phi <vscale x 2 x float> [ ..., %entry ], [ %acc.vec, %loop ] |
| 208 | // %phi.scalar = extractelement <vscale x 2 x float> %phi, i64 0 |
| 209 | // %acc = call float @llvm.vector.reduce.fadd.nxv2f32(float %x, |
| 210 | // <vscale x 2 x float> %vec) |
| 211 | // %acc.vec = insertelement <vscale x 2 x float> poison, float %acc.next, i64 0 |
| 212 | // |
| 213 | // Which eliminates the scalar -> vector -> scalar crossing during instruction |
| 214 | // selection. |
| 215 | bool RISCVCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) { |
| 216 | if (expandVPStrideLoad(I)) |
| 217 | return true; |
| 218 | |
| 219 | if (widenVPMerge(II&: I)) |
| 220 | return true; |
| 221 | |
| 222 | if (I.getIntrinsicID() != Intrinsic::vector_reduce_fadd && |
| 223 | !isa<VPReductionIntrinsic>(Val: &I)) |
| 224 | return false; |
| 225 | |
| 226 | auto *PHI = dyn_cast<PHINode>(Val: I.getOperand(i_nocapture: 0)); |
| 227 | if (!PHI || !PHI->hasOneUse() || |
| 228 | !llvm::is_contained(Range: PHI->incoming_values(), Element: &I)) |
| 229 | return false; |
| 230 | |
| 231 | Type *VecTy = I.getOperand(i_nocapture: 1)->getType(); |
| 232 | IRBuilder<> Builder(PHI); |
| 233 | auto *VecPHI = Builder.CreatePHI(Ty: VecTy, NumReservedValues: PHI->getNumIncomingValues()); |
| 234 | |
| 235 | for (auto *BB : PHI->blocks()) { |
| 236 | Builder.SetInsertPoint(BB->getTerminator()); |
| 237 | Value *InsertElt = Builder.CreateInsertElement( |
| 238 | VecTy, NewElt: PHI->getIncomingValueForBlock(BB), Idx: (uint64_t)0); |
| 239 | VecPHI->addIncoming(V: InsertElt, BB); |
| 240 | } |
| 241 | |
| 242 | Builder.SetInsertPoint(&I); |
| 243 | I.setOperand(i_nocapture: 0, Val_nocapture: Builder.CreateExtractElement(Vec: VecPHI, Idx: (uint64_t)0)); |
| 244 | |
| 245 | PHI->eraseFromParent(); |
| 246 | |
| 247 | return true; |
| 248 | } |
| 249 | |
| 250 | // Always expand zero strided loads so we match more .vx splat patterns, even if |
| 251 | // we have +optimized-zero-stride-loads. RISCVDAGToDAGISel::Select will convert |
| 252 | // it back to a strided load if it's optimized. |
| 253 | bool RISCVCodeGenPrepare::expandVPStrideLoad(IntrinsicInst &II) { |
| 254 | Value *BasePtr, *VL; |
| 255 | |
| 256 | using namespace PatternMatch; |
| 257 | if (!match(V: &II, P: m_Intrinsic<Intrinsic::experimental_vp_strided_load>( |
| 258 | Op0: m_Value(V&: BasePtr), Op1: m_Zero(), Op2: m_AllOnes(), Op3: m_Value(V&: VL)))) |
| 259 | return false; |
| 260 | |
| 261 | // If SEW>XLEN then a splat will get lowered as a zero strided load anyway, so |
| 262 | // avoid expanding here. |
| 263 | if (II.getType()->getScalarSizeInBits() > ST->getXLen()) |
| 264 | return false; |
| 265 | |
| 266 | if (!isKnownNonZero(V: VL, Q: {*DL, DT, nullptr, &II})) |
| 267 | return false; |
| 268 | |
| 269 | auto *VTy = cast<VectorType>(Val: II.getType()); |
| 270 | |
| 271 | IRBuilder<> Builder(&II); |
| 272 | Type *STy = VTy->getElementType(); |
| 273 | Value *Val = Builder.CreateLoad(Ty: STy, Ptr: BasePtr); |
| 274 | Value *Res = Builder.CreateIntrinsic( |
| 275 | ID: Intrinsic::vp_merge, Types: VTy, |
| 276 | Args: {II.getOperand(i_nocapture: 2), Builder.CreateVectorSplat(EC: VTy->getElementCount(), V: Val), |
| 277 | PoisonValue::get(T: VTy), VL}); |
| 278 | |
| 279 | II.replaceAllUsesWith(V: Res); |
| 280 | II.eraseFromParent(); |
| 281 | return true; |
| 282 | } |
| 283 | |
| 284 | bool RISCVCodeGenPrepare::run() { |
| 285 | bool MadeChange = false; |
| 286 | for (auto &BB : F) |
| 287 | for (Instruction &I : llvm::make_early_inc_range(Range&: BB)) |
| 288 | MadeChange |= visit(I); |
| 289 | |
| 290 | return MadeChange; |
| 291 | } |
| 292 | |
| 293 | bool RISCVCodeGenPrepareLegacyPass::runOnFunction(Function &F) { |
| 294 | if (skipFunction(F)) |
| 295 | return false; |
| 296 | |
| 297 | auto &TPC = getAnalysis<TargetPassConfig>(); |
| 298 | auto &TM = TPC.getTM<RISCVTargetMachine>(); |
| 299 | auto ST = &TM.getSubtarget<RISCVSubtarget>(F); |
| 300 | auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
| 301 | |
| 302 | RISCVCodeGenPrepare RVCGP(F, DT, ST); |
| 303 | return RVCGP.run(); |
| 304 | } |
| 305 | |
| 306 | INITIALIZE_PASS_BEGIN(RISCVCodeGenPrepareLegacyPass, DEBUG_TYPE, PASS_NAME, |
| 307 | false, false) |
| 308 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
| 309 | INITIALIZE_PASS_END(RISCVCodeGenPrepareLegacyPass, DEBUG_TYPE, PASS_NAME, false, |
| 310 | false) |
| 311 | |
| 312 | char RISCVCodeGenPrepareLegacyPass::ID = 0; |
| 313 | |
| 314 | FunctionPass *llvm::createRISCVCodeGenPrepareLegacyPass() { |
| 315 | return new RISCVCodeGenPrepareLegacyPass(); |
| 316 | } |
| 317 | |
| 318 | PreservedAnalyses RISCVCodeGenPreparePass::run(Function &F, |
| 319 | FunctionAnalysisManager &FAM) { |
| 320 | DominatorTree *DT = &FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 321 | auto ST = &TM->getSubtarget<RISCVSubtarget>(F); |
| 322 | bool Changed = RISCVCodeGenPrepare(F, DT, ST).run(); |
| 323 | if (!Changed) |
| 324 | return PreservedAnalyses::all(); |
| 325 | |
| 326 | PreservedAnalyses PA = PreservedAnalyses::none(); |
| 327 | PA.preserveSet<CFGAnalyses>(); |
| 328 | return PA; |
| 329 | } |
| 330 | |